What makes Spotify so brilliant is the access to so many playlists curated and updated with new music, it is the essence of music discovery like getting a mixtape from a friend. However, given the sheer breadth of playlists available it can be quite unbecoming to find new playlists and instead resort to listening to the same set of playlists. Making a radio based on a playlist can find some new songs, but the algorithm seems to favor playing music already liked disproportionately to "new music".
Perhaps this is the fault of whatever seeds the radio system with tags and artists to base the playlist on as they tend stick to a very narrow band of musical taste. I have found it quite impossible to generate a radio station that plays many different genres even if seeded with a massive 24 hour collaborative playlist my friends and I made.
Thus I suggest implementing a new "playlist generation" tool that allows users to select tags such as genre, mood, artist, decade, etc. that then uses the radio system to find related music with an emphasis on discovery over the same old songs, perhaps add a repeatability limit.
Here's a rough framework:
Create a playlist and name it
Bring the user to an interface with different categorical bins and allow them to add playlist "seeds" (genre, mood, decade, artist, etc.)
After this the radio system takes this information and produces the initial playlist
As the user listens and provides feedback the radio system doesn't endlessly play music, instead at an interval chosen by the user the playlist updates and combines the initial seeding with the feedback to produce more music (essentially a customizable discover weekly)
Users can then share these with friends and follow other user playlists